Trace-X cause attribution must carry per-bit care masks through ports

When tracing an X back through RTL by re-evaluating the active assignment with dumped values, a good generic test for which operand caused the X is counterfactual: set the operand to all 0s and to all 1s and see whether the X bits of the result shrink both times (this handles a & 0 masking, ?: merges, x select indices and arithmetic alike). But do it per bit and pass the bits that matter to the next step: otherwise qd <= in.data on a packed struct port follows the whole struct and blames an unrelated x field (in.id) several hops later. Also with VCS: VCD $dumpvars omits unpacked memories entirely while FSDB $fsdbDumpvars(0,top,"+all") dumps them as mem[0..N], so on VCD a memory read must be reconstructed by replaying the write port; VCS names static locals of unnamed always_comb blocks unnamed$$_0.<name>; and an if (x) takes the else branch in plain sim, so a dumped x-merge on such a branch means the run used -xprop.